Disclosure: I build WatermarkLift. This is an implementation note about a free, browser-local tool for the visible bottom-right Gemini star, not an independent product review. I used AI assistance to draft this post and checked its claims against the application and test records.
Start with a file, not a promise
A useful task is narrow: take an existing Gemini image you own or may edit, check whether its visible corner overlay matches a supported template, inspect the result, and download it. No account or image upload should be needed for that workflow.
Start with the original file downloaded from Gemini. A screenshot, crop or resized copy changes the evidence a template-based method relies on. A visible star is not, by itself, proof that a particular tool can safely clean the file.
The inverse equation has limits
For a known overlay, alpha blending can be written as:
observed = alpha * logo + (1 - alpha) * original
original_estimate = (observed - alpha * logo) / (1 - alpha)
That is different from asking an inpainting model to invent a patch. It is also not a guarantee of lossless restoration. The alpha template or position can be wrong; compression, rounding and clipped colors can remove information.
WatermarkLift uses the pinned MIT-licensed @pilio/gemini-watermark-remover@1.0.46, from GargantuaX/gemini-watermark-remover. The upstream method and templates credit AllenK (Kwyshell), and the JavaScript port credits Jad. I did not invent that algorithm.
Separate a candidate from an accepted result
Our application applies an additional conservative gate to the SDK output. It checks evidence and quality flags, limits the candidate region to the lower-right part of the image, and verifies that dimensions, alpha and every pixel outside the selected region remain unchanged. This reduces some risks; it cannot eliminate false positives or false negatives.
The UI distinguishes three outcomes:
- Cleaned · review: a candidate passed the checks. It still needs visual inspection.
- Original kept: there was insufficient evidence or the candidate failed a check. No cleaned download is offered.
- Could not process: the operation failed. The original remains available; retry is offered only where the file can be retried.
In an October 9 run of six public upstream regression files, two offered cleaned results and four kept the original. Those are convenience samples distributed with the algorithm, with no independently verified generation provenance or clean ground truth. Two out of six is not a real-world accuracy estimate. The guide includes the files, outcomes and limitations.
Review before downloading
Compare the original and result, then inspect the corner at 1:1. Look for a remaining star, dark halo, texture damage or color shift. If the result looks worse, keep the original rather than repeatedly processing the same file.
An edited export is a newly encoded PNG, even when the input is JPEG or WebP. Keeping the same dimensions does not restore detail absent from the input. The exact original is separately downloadable, and the ZIP includes only accepted cleaned PNGs.
The current input scope is opaque, static PNG/JPEG/WebP: up to 10 files, 20 MiB each and 100 MiB per batch; 256–4096 pixels per side, 12 MP per image and 24 MP per batch. Transparent images, animation and video are unsupported.
Keep the browser responsive and failure recoverable
Processing runs sequentially in Workers with a deadline. Cancel terminates active work and stops the queue; validated cancelled files can be retried. A failed file does not stop the remaining batch. Preview object URLs are revoked when their images are removed, and temporary download URLs expire. Memory limits still matter because decoded images are much larger than compressed files.
Automated checks cover duplicate imports, cancellation/retry, conservative negative results, failures, and PNG/original/ZIP contents. A download event means a download was initiated; it does not prove that a person saved or used the file. Mobile emulation also does not replace testing on a physical phone.
Visible cleanup is not provenance removal
This tool provides no dedicated SynthID or C2PA removal or verification. PNG re-encoding may change metadata or affect invisible signals, so keep the original. Removing a star does not establish that the image is non-AI.
For future generations, check Google’s current Media Watermark setting, whose availability depends on plan and region. That setting does not edit an existing local image.
What feedback would help
If you already have a supported Gemini image and permission to edit it, you can try WatermarkLift. I am looking for feedback on whether you can import it, understand the outcome, inspect the corner and download the intended file. You do not need to share the image: browser, file format, which step failed and the displayed status are enough to start. Please do not post private images.
Images are processed locally. Ahrefs measures visits and fixed tool events; Google Analytics loads only after consent. Neither is given image files or filenames. The privacy notice explains the distinction.
Top comments (1)
Official Platform Update
Security protocols have been updated for all developer accounts.